X-Git-Url: https://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu-qsbr.c;h=0e1252fdbba707e4515cbbabe8c8a837a6408c79;hp=fd0ebf6c4a710ae2bed98e6a919f424eb05bc992;hb=0ecb3fde04135d76545ce55d80abea9aef465b54;hpb=9f1621ca8d6ce9178343ae24aeef406c0ac2564a diff --git a/urcu-qsbr.c b/urcu-qsbr.c index fd0ebf6..0e1252f 100644 --- a/urcu-qsbr.c +++ b/urcu-qsbr.c @@ -1,7 +1,7 @@ /* - * urcu.c + * urcu-qsbr.c * - * Userspace RCU library + * Userspace RCU QSBR library * * Copyright (c) 2009 Mathieu Desnoyers * Copyright (c) 2009 Paul E. McKenney, IBM Corporation. @@ -32,205 +32,243 @@ #include #include -#include "urcu-static.h" +#define BUILD_QSBR_LIB +#include "urcu-qsbr-static.h" /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ -#include "urcu.h" +#include "urcu-qsbr.h" -void __attribute__((constructor)) urcu_init(void); -void __attribute__((destructor)) urcu_exit(void); +void __attribute__((destructor)) rcu_exit(void); -int init_done; +static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; -pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER; +int gp_futex; /* * Global grace period counter. */ -long urcu_gp_ctr = 0; +unsigned long rcu_gp_ctr = RCU_GP_ONLINE; /* * Written to only by each individual reader. Read by both the reader and the * writers. */ -long __thread urcu_active_readers; - -/* Thread IDs of registered readers */ -#define INIT_NUM_THREADS 4 - -struct reader_registry { - pthread_t tid; - long *urcu_active_readers; - char *need_mb; -}; +struct rcu_reader __thread rcu_reader; #ifdef DEBUG_YIELD unsigned int yield_active; unsigned int __thread rand_yield; #endif -static struct reader_registry *registry; -static char __thread need_mb; -static int num_readers, alloc_readers; +static LIST_HEAD(registry); -void internal_urcu_lock(void) +static void mutex_lock(pthread_mutex_t *mutex) { int ret; #ifndef DISTRUST_SIGNALS_EXTREME - ret = pthread_mutex_lock(&urcu_mutex); + ret = pthread_mutex_lock(mutex); if (ret) { perror("Error in pthread mutex lock"); exit(-1); } #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ - while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) { + while ((ret = pthread_mutex_trylock(mutex)) != 0) { if (ret != EBUSY && ret != EINTR) { printf("ret = %d, errno = %d\n", ret, errno); perror("Error in pthread mutex lock"); exit(-1); } - if (need_mb) { - smp_mb(); - need_mb = 0; - smp_mb(); - } poll(NULL,0,10); } #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ } -void internal_urcu_unlock(void) +static void mutex_unlock(pthread_mutex_t *mutex) { int ret; - ret = pthread_mutex_unlock(&urcu_mutex); + ret = pthread_mutex_unlock(mutex); if (ret) { perror("Error in pthread mutex unlock"); exit(-1); } } -#ifdef DEBUG_FULL_MB -#ifdef HAS_INCOHERENT_CACHES -static void force_mb_single_thread(struct reader_registry *index) +/* + * synchronize_rcu() waiting. Single thread. + */ +static void wait_gp(void) { - smp_mb(); + /* Read reader_gp before read futex */ + smp_rmb(); + if (uatomic_read(&gp_futex) == -1) + futex_noasync(&gp_futex, FUTEX_WAIT, -1, + NULL, NULL, 0); } -#endif /* #ifdef HAS_INCOHERENT_CACHES */ -static void force_mb_all_threads(void) -{ - smp_mb(); -} -#else /* #ifdef DEBUG_FULL_MB */ -#ifdef HAS_INCOHERENT_CACHES -static void force_mb_single_thread(struct reader_registry *index) +static void update_counter_and_wait(void) { - assert(registry); + LIST_HEAD(qsreaders); + int wait_loops = 0; + struct rcu_reader *index, *tmp; + +#if (BITS_PER_LONG < 64) + /* Switch parity: 0 -> 1, 1 -> 0 */ + STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR); +#else /* !(BITS_PER_LONG < 64) */ + /* Increment current G.P. */ + STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR); +#endif /* !(BITS_PER_LONG < 64) */ + /* - * pthread_kill has a smp_mb(). But beware, we assume it performs - * a cache flush on architectures with non-coherent cache. Let's play - * safe and don't assume anything : we use smp_mc() to make sure the - * cache flush is enforced. + * Must commit rcu_gp_ctr update to memory before waiting for quiescent + * state. Failure to do so could result in the writer waiting forever + * while new readers are always accessing data (no progress). Enforce + * compiler-order of store to rcu_gp_ctr before load rcu_reader ctr. + */ + barrier(); + + /* + * Adding a smp_mb() which is _not_ formally required, but makes the + * model easier to understand. It does not have a big performance impact + * anyway, given this is the write-side. */ - *index->need_mb = 1; - smp_mc(); /* write ->need_mb before sending the signals */ - pthread_kill(index->tid, SIGURCU); smp_mb(); + /* - * Wait for sighandler (and thus mb()) to execute on every thread. - * BUSY-LOOP. + * Wait for each thread rcu_reader_qs_gp count to become 0. */ - while (*index->need_mb) { - poll(NULL, 0, 1); + for (;;) { + wait_loops++; + if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { + uatomic_dec(&gp_futex); + /* Write futex before read reader_gp */ + smp_mb(); + } + + list_for_each_entry_safe(index, tmp, ®istry, node) { + if (!rcu_gp_ongoing(&index->ctr)) + list_move(&index->node, &qsreaders); + } + + if (list_empty(®istry)) { + if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { + /* Read reader_gp before write futex */ + smp_mb(); + uatomic_set(&gp_futex, 0); + } + break; + } else { + if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { + wait_gp(); + } else { +#ifndef HAS_INCOHERENT_CACHES + cpu_relax(); +#else /* #ifndef HAS_INCOHERENT_CACHES */ + smp_mb(); +#endif /* #else #ifndef HAS_INCOHERENT_CACHES */ + } + } } - smp_mb(); /* read ->need_mb before ending the barrier */ + /* put back the reader list in the registry */ + list_splice(&qsreaders, ®istry); } -#endif /* #ifdef HAS_INCOHERENT_CACHES */ -static void force_mb_all_threads(void) +/* + * Using a two-subphases algorithm for architectures with smaller than 64-bit + * long-size to ensure we do not encounter an overflow bug. + */ + +#if (BITS_PER_LONG < 64) +void synchronize_rcu(void) { - struct reader_registry *index; + unsigned long was_online; + + was_online = rcu_reader.ctr; + + /* All threads should read qparity before accessing data structure + * where new ptr points to. + */ + /* Write new ptr before changing the qparity */ + smp_mb(); + /* - * Ask for each threads to execute a smp_mb() so we can consider the - * compiler barriers around rcu read lock as real memory barriers. + * Mark the writer thread offline to make sure we don't wait for + * our own quiescent state. This allows using synchronize_rcu() in + * threads registered as readers. */ - if (!registry) - return; + if (was_online) + STORE_SHARED(rcu_reader.ctr, 0); + + mutex_lock(&rcu_gp_lock); + + if (list_empty(®istry)) + goto out; + /* - * pthread_kill has a smp_mb(). But beware, we assume it performs - * a cache flush on architectures with non-coherent cache. Let's play - * safe and don't assume anything : we use smp_mc() to make sure the - * cache flush is enforced. + * Wait for previous parity to be empty of readers. */ - for (index = registry; index < registry + num_readers; index++) { - *index->need_mb = 1; - smp_mc(); /* write need_mb before sending the signal */ - pthread_kill(index->tid, SIGURCU); - } + update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */ + /* - * Wait for sighandler (and thus mb()) to execute on every thread. - * - * Note that the pthread_kill() will never be executed on systems - * that correctly deliver signals in a timely manner. However, it - * is not uncommon for kernels to have bugs that can result in - * lost or unduly delayed signals. - * - * If you are seeing the below pthread_kill() executing much at - * all, we suggest testing the underlying kernel and filing the - * relevant bug report. For Linux kernels, we recommend getting - * the Linux Test Project (LTP). + * Must finish waiting for quiescent state for parity 0 before + * committing next rcu_gp_ctr update to memory. Failure to do so could + * result in the writer waiting forever while new readers are always + * accessing data (no progress). Enforce compiler-order of load + * rcu_reader ctr before store to rcu_gp_ctr. */ - for (index = registry; index < registry + num_readers; index++) { - while (*index->need_mb) { - pthread_kill(index->tid, SIGURCU); - poll(NULL, 0, 1); - } - } - smp_mb(); /* read ->need_mb before ending the barrier */ -} -#endif /* #else #ifdef DEBUG_FULL_MB */ + barrier(); -void wait_for_quiescent_state(void) -{ - struct reader_registry *index; + /* + * Adding a smp_mb() which is _not_ formally required, but makes the + * model easier to understand. It does not have a big performance impact + * anyway, given this is the write-side. + */ + smp_mb(); - if (!registry) - return; /* - * Wait for each thread urcu_active_readers count to become 0. + * Wait for previous parity to be empty of readers. */ - for (index = registry; index < registry + num_readers; index++) { -#ifndef HAS_INCOHERENT_CACHES - while (rcu_old_gp_ongoing(index->urcu_active_readers)) - cpu_relax(); -#else /* #ifndef HAS_INCOHERENT_CACHES */ - int wait_loops = 0; - /* - * BUSY-LOOP. Force the reader thread to commit its - * urcu_active_readers update to memory if we wait for too long. - */ - while (rcu_old_gp_ongoing(index->urcu_active_readers)) { - if (wait_loops++ == KICK_READER_LOOPS) { - force_mb_single_thread(index); - wait_loops = 0; - } else { - cpu_relax(); - } - } -#endif /* #else #ifndef HAS_INCOHERENT_CACHES */ - } -} + update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */ +out: + mutex_unlock(&rcu_gp_lock); + /* + * Finish waiting for reader threads before letting the old ptr being + * freed. + */ + if (was_online) + _STORE_SHARED(rcu_reader.ctr, LOAD_SHARED(rcu_gp_ctr)); + smp_mb(); +} +#else /* !(BITS_PER_LONG < 64) */ void synchronize_rcu(void) { - internal_urcu_lock(); - force_mb_all_threads(); - urcu_gp_ctr += 2; - wait_for_quiescent_state(); - force_mb_all_threads(); - internal_urcu_unlock(); + unsigned long was_online; + + was_online = rcu_reader.ctr; + + /* + * Mark the writer thread offline to make sure we don't wait for + * our own quiescent state. This allows using synchronize_rcu() in + * threads registered as readers. + */ + smp_mb(); + if (was_online) + STORE_SHARED(rcu_reader.ctr, 0); + + mutex_lock(&rcu_gp_lock); + if (list_empty(®istry)) + goto out; + update_counter_and_wait(); +out: + mutex_unlock(&rcu_gp_lock); + + if (was_online) + _STORE_SHARED(rcu_reader.ctr, LOAD_SHARED(rcu_gp_ctr)); + smp_mb(); } +#endif /* !(BITS_PER_LONG < 64) */ /* * library wrappers to be used by non-LGPL compatible source code. @@ -246,145 +284,45 @@ void rcu_read_unlock(void) _rcu_read_unlock(); } -void *rcu_dereference(void *p) -{ - return _rcu_dereference(p); -} - -void *rcu_assign_pointer_sym(void **p, void *v) -{ - wmb(); - return STORE_SHARED(p, v); -} - -void *rcu_xchg_pointer_sym(void **p, void *v) +void rcu_quiescent_state(void) { - wmb(); - return xchg(p, v); + _rcu_quiescent_state(); } -void *rcu_publish_content_sym(void **p, void *v) +void rcu_thread_offline(void) { - void *oldptr; - - oldptr = _rcu_xchg_pointer(p, v); - synchronize_rcu(); - return oldptr; -} - -static void rcu_add_reader(pthread_t id) -{ - struct reader_registry *oldarray; - - if (!registry) { - alloc_readers = INIT_NUM_THREADS; - num_readers = 0; - registry = - malloc(sizeof(struct reader_registry) * alloc_readers); - } - if (alloc_readers < num_readers + 1) { - oldarray = registry; - registry = malloc(sizeof(struct reader_registry) - * (alloc_readers << 1)); - memcpy(registry, oldarray, - sizeof(struct reader_registry) * alloc_readers); - alloc_readers <<= 1; - free(oldarray); - } - registry[num_readers].tid = id; - /* reference to the TLS of _this_ reader thread. */ - registry[num_readers].urcu_active_readers = &urcu_active_readers; - registry[num_readers].need_mb = &need_mb; - num_readers++; + _rcu_thread_offline(); } -/* - * Never shrink (implementation limitation). - * This is O(nb threads). Eventually use a hash table. - */ -static void rcu_remove_reader(pthread_t id) +void rcu_thread_online(void) { - struct reader_registry *index; - - assert(registry != NULL); - for (index = registry; index < registry + num_readers; index++) { - if (pthread_equal(index->tid, id)) { - memcpy(index, ®istry[num_readers - 1], - sizeof(struct reader_registry)); - registry[num_readers - 1].tid = 0; - registry[num_readers - 1].urcu_active_readers = NULL; - num_readers--; - return; - } - } - /* Hrm not found, forgot to register ? */ - assert(0); + _rcu_thread_online(); } void rcu_register_thread(void) { - internal_urcu_lock(); - urcu_init(); /* In case gcc does not support constructor attribute */ - rcu_add_reader(pthread_self()); - internal_urcu_unlock(); -} + rcu_reader.tid = pthread_self(); + assert(rcu_reader.ctr == 0); -void rcu_unregister_thread(void) -{ - internal_urcu_lock(); - rcu_remove_reader(pthread_self()); - internal_urcu_unlock(); + mutex_lock(&rcu_gp_lock); + list_add(&rcu_reader.node, ®istry); + mutex_unlock(&rcu_gp_lock); + _rcu_thread_online(); } -#ifndef DEBUG_FULL_MB -static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context) +void rcu_unregister_thread(void) { /* - * Executing this smp_mb() is the only purpose of this signal handler. - * It punctually promotes barrier() into smp_mb() on every thread it is - * executed on. + * We have to make the thread offline otherwise we end up dealocking + * with a waiting writer. */ - smp_mb(); - need_mb = 0; - smp_mb(); + _rcu_thread_offline(); + mutex_lock(&rcu_gp_lock); + list_del(&rcu_reader.node); + mutex_unlock(&rcu_gp_lock); } -/* - * urcu_init constructor. Called when the library is linked, but also when - * reader threads are calling rcu_register_thread(). - * Should only be called by a single thread at a given time. This is ensured by - * holing the internal_urcu_lock() from rcu_register_thread() or by running at - * library load time, which should not be executed by multiple threads nor - * concurrently with rcu_register_thread() anyway. - */ -void urcu_init(void) +void rcu_exit(void) { - struct sigaction act; - int ret; - - if (init_done) - return; - init_done = 1; - - act.sa_sigaction = sigurcu_handler; - ret = sigaction(SIGURCU, &act, NULL); - if (ret) { - perror("Error in sigaction"); - exit(-1); - } -} - -void urcu_exit(void) -{ - struct sigaction act; - int ret; - - ret = sigaction(SIGURCU, NULL, &act); - if (ret) { - perror("Error in sigaction"); - exit(-1); - } - assert(act.sa_sigaction == sigurcu_handler); - free(registry); + assert(list_empty(®istry)); } -#endif /* #ifndef DEBUG_FULL_MB */